Associations Between Perceived Generative AI Support Features and Self-Reported Deep Knowledge Construction Among Chinese University Students

As generative AI becomes increasingly embedded in higher education, a central sustainability challenge is whether greater access to AI support can be achieved without weakening the learner’s self-regulation required for meaningful knowledge construction. Yet it remains unclear which forms of perceived AI support are associated with deeper knowledge construction and through which learner-level regulatory pathways this association operates. This study examines six perceived support features—fine-grainedness (FG), transparency in reasoning (TR), adaptivity in explanation (AE), strategy instruction embeddedness (SIE), multimodal richness (MR), and social presence (SP)—and their associations with self-reported deep knowledge construction (DKC). Survey data from 626 Chinese university students were analysed using covariance-based structural equation modelling, latent-variable bootstrap analysis, and fuzzy-set qualitative comparative analysis. FG, TR, and AE showed significant direct associations with DKC, whereas SIE, MR, and SP did not. Ten of 12 hypothesised indirect associations involving self-efficacy (SE) and metacognitive monitoring (MM) were supported; the AE → MM → DKC and MR → MM → DKC pathways were not. Two configurations were associated with high DKC, with TR and AE recurring as core conditions. Given the cross-sectional, self-reported design, these findings indicate associations rather than causal effects or direct sustainability outcomes. They inform the educationally sustainable integration of generative AI by identifying learner-regulatory processes associated with knowledge construction.